Supporting irregular distributions using data-parallel languages

R. Ponnusamy, Yuan-Shin Hwang, R. Das, J.H. Saltz, A. Choudhary, G. Fox
1995 IEEE Parallel & Distributed Technology Systems & Applications  
g Languages such as Fortran D provide irregular distribution schemes that can efficiently support irregular-problems. Irregular distributions can also be emulated in HPF. Compilers can incorporate runtime procedures to automatically support these distributions. 12 0 n distributed-memory machines, large data arrays need to be partitioned between local processor memories. These partitioned data arrays are called distributed arrays. Many applications can be efficiently implemented by using simple
more » ... chemes for mapping distributed arrays. One such scheme is BLOCK distribution, which divides an array into contiguous, equal-sized subarrays and assigns each subarray to a different processor. Another is CYCLIC distribution, which assigns consecutively indexed array elements to processors in round-robin fashion. However, more complex distributions are required to efficiently execute in-egularpl-oblem such as computational fluid dynamics codes, molecular dynamics codes, diagonal or polynomial preconditioned iterative linear solvers, and time-dependent flame-modeling codes. Researchers have developed a variety of methods to obtain data mappings that optimize the communication requirements of irregular problems. 1-3 These methods produce irregular distributions. The Fortran D," Fortran 90D, and Vienna Fortranj data-parallel languages support irregular data distributions. Fortran D and Fortran 90D let a programmer explicitly specify an irregular distribution using an array, to specify a mapping of array elements to processors. (Fortran D is Fortran 77 with data distribution; Fortran 90D is Fortran 90 with data distribution.) Vienna Fortran lets developers define functions to describe irregular distributions. However, the current version of High Performance Fortran does not directly support irregular distributions6 Also, in irregular problems, data-access patterns and workload are usu-1063-655?/95/$1,00 0 1995 IEEE. IEEE Parallel & Distributed Technology
doi:10.1109/88.384581 fatcat:qscb4olexveehlp7swtnbm3ymm